4 ms·

And these are high end 'prosumer' parts. Running the CPU at peak 80-100% load for a significant amount of time should be considered the intended application.

ianhowson 9y ago
What does 80-100% even mean, though? The CPU cores are merely one component of the CPU 'package'. The GPU on most Intel parts consumes way more power than the CPU cores when running flat-out. Even just within the CPU cores, different ops consume vastly different amount of energy while still 'maxing out' utilisation. And we haven't even considered accelerator hardware within the CPU which is totally independent of the CPU cores but can still consume a large amount of energy (e.g. H.264, AES). Intel no doubt calculate a maximum power consumption figure where every bit of hardware on the CPU is consuming as much as it can. It'd be utterly meaningless -- probably on the order of 100W for a 2-4 core laptop part or 400W for a server part. You can design a computer to handle those figures, but you don't want to. So TDP reflects intended usage rather than peak demand. If you're a 'prosumer', you're presumably smart enough to size your power supply and cooling appropriately. TDP is not the number you're looking for.
- Coding_Cat 9y ago
  Ah yes fair point, my personal bias is showing a bit. I would consider numerics (AVX512 or AVX2) to be the main interest, and it is often what the type of workstations I'm familiar with need. It also tends to consume the most power when run at full load by a significant margin. the 80-100% is bit of a vague number, I was just kinda referring to the cpu utilization as reported by the OS. For numerics I would consider a realistic full load 75%+ pipeline fill time on the AVX codepaths. With a lot of compute you can get quite close to 100% quite quickly if your software is highly optimized (and uses enough compute to hide the memory loads). TDP should just be 100% pipelined load of the highest consuming codepath. But my general point is still that these benchmarks should not exceed the TDP for these types of parts. For everyday use these loads might not be realistic. But for a lot of people looking to spend $2000(!) on a single CPU, these benchmarks are indicative of real life usage.

ianhowson 9y ago

What does 80-100% even mean, though? The CPU cores are merely one component of the CPU 'package'. The GPU on most Intel parts consumes way more power than the CPU cores when running flat-out. Even just within the CPU cores, different ops consume vastly different amount of energy while still 'maxing out' utilisation. And we haven't even considered accelerator hardware within the CPU which is totally independent of the CPU cores but can still consume a large amount of energy (e.g. H.264, AES). Intel no doubt calculate a maximum power consumption figure where every bit of hardware on the CPU is consuming as much as it can. It'd be utterly meaningless -- probably on the order of 100W for a 2-4 core laptop part or 400W for a server part. You can design a computer to handle those figures, but you don't want to. So TDP reflects intended usage rather than peak demand. If you're a 'prosumer', you're presumably smart enough to size your power supply and cooling appropriately. TDP is not the number you're looking for.

Coding_Cat 9y ago
Ah yes fair point, my personal bias is showing a bit. I would consider numerics (AVX512 or AVX2) to be the main interest, and it is often what the type of workstations I'm familiar with need. It also tends to consume the most power when run at full load by a significant margin. the 80-100% is bit of a vague number, I was just kinda referring to the cpu utilization as reported by the OS. For numerics I would consider a realistic full load 75%+ pipeline fill time on the AVX codepaths. With a lot of compute you can get quite close to 100% quite quickly if your software is highly optimized (and uses enough compute to hide the memory loads). TDP should just be 100% pipelined load of the highest consuming codepath. But my general point is still that these benchmarks should not exceed the TDP for these types of parts. For everyday use these loads might not be realistic. But for a lot of people looking to spend $2000(!) on a single CPU, these benchmarks are indicative of real life usage.